Probabilistic Estimation of the Gaze Region of the Driver using Dense Classification
Authors:
- 1 November 2018
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 697-702
Abstract
The ability to monitor the visual attention of a driver is a useful feature for smart vehicles to understand the driver's intents and behaviors. The gaze angle of the driver is not deterministically related to his/her head pose due to the interplay between head and eye movements. Therefore, this study aims to establish a probabilistic relationship using deep learning. While probabilistic regression techniques such as Gaussian process regression (GPR) has been previously used to predict the visual attention of a driver, the proposed deep learning framework is a more generic approach that does not make assumptions, learning the relationship between gaze and head pose from the data. In our formulation, the continuous gaze angles are converted into intervals and the grid of the quantized angles is treated as an image for dense prediction. We rely on convolutional neural networks (CNNs) with upsampling to map the six degrees of freedom of the orientation and position of the head into gaze angles. We train and evaluate the proposed network with data collected from drivers who were asked to look at predetermined locations inside a car during naturalistic driving recordings. The proposed model obtains very promising results, where the size of the gaze region with 95% accuracy is only 11.73% of a half sphere centered at the driver, which approximates his/her field of view. The architecture offers an appealing and general solution to convert regression problems into dense classification problems.Keywords
This publication has 20 references indexed in Scilit:
- Analyzing the relationship between head pose and gaze to model driver visual attention2016
- Learning Deconvolution Network for Semantic Segmentation2015
- Detecting Drivers' Mirror-Checking Actions and Its Application to Maneuver and Secondary Task RecognitionIEEE Transactions on Intelligent Transportation Systems, 2015
- Fully convolutional networks for semantic segmentation2015
- Predicting Perceived Visual and Cognitive Distractions of Drivers With Multimodal FeaturesIEEE Transactions on Intelligent Transportation Systems, 2014
- Robust and continuous estimation of driver gaze zone by dynamic analysis of multiple face videos2014
- Adaptive deconvolutional networks for mid and high level feature learning2011
- Combining cognitive and visual distraction: Less than the sum of its partsAccident Analysis & Prevention, 2010
- A Survey on Transfer LearningIEEE Transactions on Knowledge and Data Engineering, 2009
- A Simple Approach to Ordinal Classification2001